PRISMA
The Challenge
Name:
Open Challenge
Domain:
Autonomous maritime inspection and predictive maintenance: historical defect evolution modelling, domain-adaptive deep learning, digital twin integration and decision support for structural condition assessment of vessels.
Challenge proposer:
Objectives:
To add a predictive intelligence layer to the AUTOASSESS autonomous UAS inspection system so that missions can be targeted rather than exhaustive. Specifically: model historical defect evolution from longitudinal, vessel-specific inspection data; apply domain-adaptive deep learning that generalises across vessel types; integrate the predictions into the AUTOASSESS Digital Twin data environment (Cognite Data Fusion); and deliver decision-support outputs that direct drones to the highest-risk structural zones, cutting dry-dock inspection time from 15 days towards 3.
The Solution
Name:
PRISMA – Predictive Risk Intelligence for Structural Maritime Assessment
PRISMA (Predictive Risk Intelligence for Structural Maritime Assessment) adds a predictive intelligence layer to the AUTOASSESS autonomous inspection system. Autonomous drones fitted with NDT sensors scan the vast, GNSS-denied interiors of ballast tanks and cargo holds, but without prior knowledge of where structural degradation is most likely they must scan indiscriminately and exhaust scarce battery endurance on sound surfaces. PRISMA introduces Targeted Autonomous Navigation: a predictive Digital Twin driven by a Heterogeneous Graph Transformer trained on 10 to 15 years of vessel-specific Ultrasonic Thickness Measurements, visual survey logs and operational profiles. From this history it generates dynamic 3D Risk Heatmaps that ingest into the AUTOASSESS Digital Twin data environment (Cognite Data Fusion) and direct UAS mission planners to the statistically highest-risk zones before launch. By telling the robot where to look, PRISMA is the enabling component that makes it possible to compress a traditional 15-day dry-dock inspection to as little as 3 days.
PRISMA turns the historical record of a vessel into forward-looking targeting intelligence for autonomous inspection. Classification rules already require that areas of substantial corrosion found at previous surveys are re-measured, and that measurement locations reflect a vessel’s cargo and ballast history (IACS UR Z10.1 and Z7.1), yet these judgments are still made manually by an experienced surveyor working from memory and paper records. PRISMA formalises and operationalises at machine scale what the rules already require.
The core model is a Heterogeneous Graph Transformer (HGT) that represents a vessel’s structure as a graph: nodes are transverse frames, longitudinal stiffeners, brackets, welds and plating panels, and edges encode geometric adjacency, load-path coupling and corrosion-boundary proximity. By attending to these heterogeneous relationships the model learns structurally aware propagation patterns, for example how corrosion at a frame-longitudinal intersection tends to spread along the longitudinal before affecting adjacent frames. Predictive confidence is quantified with Monte Carlo Dropout, so every heatmap separates high-confidence high-risk zones (direct UAS targeting) from high-uncertainty zones (inspect to reduce uncertainty). SHAP attribution gives surveyors and class-society reviewers a human-readable rationale for each flagged zone, and Maximum Mean Discrepancy domain adaptation lets the model generalise to a new vessel type from a single complete inspection cycle, keeping the adoption barrier low.
The outputs are dynamic 3D Risk Heatmaps that ingest into the AUTOASSESS Digital Twin data environment (Cognite Data Fusion) through a formally defined data interface and become structured prioritisation inputs for the AUTOASSESS UI-DSS and mission-planning layer. PRISMA does not duplicate AUTOASSESS work; it supplies the upstream predictive input that lets the downstream mission-execution layer operate at peak efficiency.
PRISMA enters the project at TRL 5, based on prior offline validation of the HGT architecture against a 12-year archive of UTM and visual records and field tests in the Piraeus Ship Repair Zone. Over nine months it advances to TRL 7 through integration with AUTOASSESS and a live UAS inspection mission on an operational vessel at the Port of Piraeus. All core components (HGT, MC Dropout, MMD, SHAP) are published and implementable in PyTorch Geometric, so no fundamental new research is required; the innovation lies in their combination, the AUTOASSESS integration, and the step from analytical validation to operational execution.
Target results include a validated predictive Digital Twin module, a Digital Twin-integrated risk heatmap module, a defect-prediction F1-score of at least 70%, at least 60% of confirmed defects captured within the top 15% risk zones, and at least a 40% reduction in required drone scanning area relative to blind exploration. At fleet scale this supports the AUTOASSESS goal of cutting dry-dock inspection from 15 days to 3, an estimated saving of about EUR 800,000 per vessel per cycle.
The Solution Provider
bSpoke Solutions P.C. is a Greek technology company that builds intelligent systems combining artificial intelligence, automation and advanced analytics for organisations in maritime, agrifood, energy, logistics, e-health and transport. Positioning itself as a one-stop shop for digital transformation, bSpoke focuses on high-innovation, high-complexity projects that need extensive research and domain expertise, ranging from predictive maintenance and computer vision to XR-based remote ship inspection. The company pairs machine-learning and systems-integration capability and has a track record of supporting technology adoption and commercialisation in EU-funded research and innovation projects as well as private projects. Within PRISMA, bSpoke leads the full development, integration and demonstration of the predictive inspection solution.
